Machine Learning for Marine Insurance - the time is now

An explosion of data

It is no secret that the marine industry is becoming better connected and this, along with the growth in high-performance computing, means that insurers can gain access to more and more accurate data. Data like vessel movements, the state of vessel machinery and port data.

All of this information can be used on a real-time basis to gain a deeper understanding of the risk profile of fleets and vessels and therefore the appropriate insurance prices and products to offer. This deeper understanding is exactly what marine insurers need to carve out profits in an extremely challenging market.

The question is, can all of this new data be analysed and interpreted in a meaningful, and regular, way using the traditional methods? The simple answer is no.

How risk analysis and pricing is currently done

Underwriters rely on actuaries for accurate pricing models. Actuaries build pricing models using statistical models which analyse a small number of static variables, or ‘features’, such as historical losses, tonnage and flag.

One or two actuaries will work together to run a couple of different features through up to one hundred statistical iterations. This will help them segment their portfolio to come up with a new pricing model, or optimise their current model, in the hope of generating more profit.

This generally happens once a year, or sometimes far less, and can take up to six months to complete.

So, how would the current process stand up to all of this new (and very useful) dynamic, and often complex, data?  The simple answer is it wouldn’t, but Machine Learning would.

What is Machine Learning and what strategy should be adopted?

Machine Learning is the process of computers self-learning complex rules based on input data. An easy example of this is Netflix. Netflix employs machine learning to recommend new things to watch based on your viewing behaviour, the initial preferences you provided when you signed up and behavioural data collected from all of the other 130 million Netflix subscribers.

Machine learning is one stage in the Data Science workflow:

1.  Definition of the business problem: both the task to perform and definitions of success;

2.  Data acquisition;

3.  Data preparation (data cleansing, feature engineering);

4.  Machine Learning modelling;

5.  Deployment into production

For a successful Machine Learning project it’s really important to follow all the above steps as well as:

• Be patient. This is an experimental process where the optimum result is generally found after several iterations

• Focus on the data preparation stage, even if it is time consuming

Machine Learning does not perform miracles but, with the right data input and collaboration with human business expertise, the output can be rich and lead to much better and quicker decision making than traditional approaches.

How things could look

In comparison to today’s manual and laborious process, underwriters and actuaries could work together using machine learning. They could run historic, static and new behavioural data through tens, if not hundreds of models, and hundreds of thousands of iterations.

Rather than once a year, or less, they could do this on a regular, even daily basis, if they wanted to, at the push of a button.

Not only will richer data, run through more models and more iterations, lead to far greater accuracy in risk assessment and pricing, the ability to carry this out on a much more regular basis could open up significant new opportunities. Think about the changes that could happen throughout the year that might impact the risk profile of a portfolio - the operator might change, there could be seasonal impacts to cargo, new sanctions might be introduced, or high risk zones change. The ability for an insurer to dynamically change pricing and policies as they need to would no doubt have a significant impact on profit margins.

The time is now

From healthcare to transportation, Machine Learning has already improved the way entire industries operate. With access to swathes of new data and numerous business challenges to focus on, Marine Insurers now find themselves in the perfect position to leverage this new approach to follow suit. Machine Learning will help Marine Insurers achieve the accurate pricing and tailored policies that are crucial to remaining competitive in this challenging and unprofitable environment.

Authors

Dorian Lacaisse - Senior Data Scientist, Concirrus

Dorian is a Senior Data Scientist at Concirrus, using data to uncover hidden patterns that influence marine insurance risk. Having already pursued a successful global career in industrial engineering, Dorian re-trained as a Data Scientist to chase his passion for Deep Learning. Dorian predominantly works with Python and Spark in Concirrus’ Hadoop environment.

Matthew Madahar - Pre Sales and Support Manager, Concirrus

Formerly a Data Analyst at Axis, and Actuarial Technician at Munich Re, Matthew has a deep understanding of Marine Insurance Data. Matthew is passionate about the impact that data and technology will have on the marine insurance industry and, as Concirrus’ Pre Sales and Support Manager, Matthew works with Concirrus’ clients to help them unlock the value of big data and machine learning.